This page organizes papers related to artificial intelligence published around the world. This page is summarized using Google Gemini and is operated on a non-profit basis. The copyright of the paper belongs to the author and the relevant institution. When sharing, simply cite the source.
This paper discusses recent advances in self-improvement, which improves the output of large-scale language models (LLMs) through iterative refinement. To overcome the limitations of existing self-improvement methods, we propose ProActive Self-Refinement (PASR), a novel method that allows LLMs to improve their output during the generation process. PASR actively determines when and how to improve based on the model's internal state and evolving context. Extensive experiments on ten diverse tasks demonstrate that PASR significantly improves problem-solving performance. Specifically, on the Qwen3-8B model, PASR reduces average token consumption by 41.6% compared to standard generation methods while improving accuracy by 8.2%.
Takeaways, Limitations
•
Takeaways:
◦
Proposing an active self-improvement method to improve the output quality of LLM.
◦
Simultaneously reducing token consumption and improving accuracy
◦
Performance verification for various tasks
◦
Code and baselines available via GitHub
•
Limitations:
◦
Lack of details about the specific methodology (e.g., how the “internal state of the model” is utilized)
◦
Further research is needed to determine the generalizability of the proposed method.
◦
Further experiments are needed to determine whether the effects of PASR are consistent across different models and tasks.
